A Toronka*, M Defreitas, B Konkel, M Nedrud, I Zaki, A Valentine, S Cubberley, F Yin, M Bashir, K Lafata, Duke University, Durham, NC

• 
  A Toronka

• 
  
• 
  
• 
  
• 
  
• 
  
• 
  
• 
  

WE-IePD-TRACK 5-2 (Wednesday, 7/28/2021) 12:30 PM - 1:00 PM [Eastern Time (GMT-4)]

Purpose: To investigate associations between demographic, clinical, and CT-based estimates of muscle and fat mass with treatment tolerance in patients with gastric or esophageal cancer.

Methods: We retrospectively identified 142 patients with gastroesophageal adenocarcinoma treated with neoadjuvant chemotherapy+/-radiation. Clinical endpoints were defined as: treatment breaks in neoadjuvant therapy; reduction in neoadjuvant therapy dose; admission to the Emergency Department (ED) / inpatient encounters; and neoadjuvant treatment completion. Demographic and clinical features included age, gender, race, ethnicity, BMI, smoking status, pack years, TNM Stage, and treatment site. A deep learning model using the UNET architecture was developed to segment visceral fat (VF), subcutaneous (SF), and muscle on CT images from T12-L5. These values were used to derive: average transverse area of VF, SF, and muscle; ratio of VF/SF area; ratio of muscle area to total fat (VF+SF) area; skeletal muscle index (muscle area/[patient height]2). Univariate statistical associations between each clinical endpoint and the demographic, clinical, and imaging features were analyzed using T-tests or Chi-squared tests, as appropriate, with Bonferroni correction. Multivariate machine learning models were then developed for each clinical endpoint, where Regularized Logistic Regression, Support Vector Machine (SVM), and Random Forest (RF) formalisms were compared. Model training and hyperparameter tuning was achieved with 5-fold cross-validation, and model generalization was evaluated based on a stratified Monte Carlo sub-sampling technique.

Results: On univariate analysis, age was associated with breaks in neoadjuvant therapy (p=0.047), and VF/SF was associated with ED admissions (p=0.023). On multivariate analysis, SVM modeling achieved the best performance regarding breaks in neoadjuvant therapy (AUC=0.64) and neoadjuvant treatment completion (AUC=0.64). In contrast, RF modeling achieved the best performance for ED admissions (AUC=0.64) and dose reductions (AUC=0.54).

Conclusion: Our preliminary analysis suggests that demographic, clinical, and imaging features may be associated with adverse outcome in patients with gastric and/or esophageal cancer.

ePosters

Keywords

CT, Statistical Analysis, Modeling

Taxonomy

IM- CT: Machine learning, computer vision

Contact Email